Abstract: A novel two-step dictionary learning (DL) framework was proposed to dynamically adjust the overcomplete basis (a.k.a. dictionary) for matching the changes of the RSS measurements, and then the sparse solution can better represent location estimations. Moreover, a modified re-weighting l1 norm minimization algorithm was proposed to improve reconstruction performance for sparse signals. The effectiveness of the proposed scheme is demonstrated by experimental results where the locations of targets can be obtained from noisy signals, even if the number of targets is not known a priori.